Food web interactions govern how ecosystems respond to climate change and biodiversity loss. Modularity, where subgroups of species interact more often with each other than with species outside their subgroup, is a key structural feature which has been linked to food web stability. We sought to address the lack of understanding of how modularity varies among ecosystems by comparing the structure of four highly resolved marine food webs, using a simulated annealing algorithm to identify network modules and Random Forest models to predict the distribution of species across modules based on a set of eight functional traits. Modules in two offshore networks were partitioned largely by trophic level, creating an interdependence among them, whereas modules in two semi-enclosed bays were generally separated into energy channels with less trophic separation and containing distinct basal resources, providing greater redundancy in the flow of energy through the network. Foraging habitat and mobility predicted module membership in all networks, whilst body mass and foraging strategy also differentiated modules in the offshore and bay ecosystems, respectively. Environmental heterogeneity may be a key factor driving the differences in modularity and the relative importance of functional traits for predicting module membership. Our results indicate that, in addition to overall network modularity, the trophic structure of modules within food webs should be considered when making inferences about ecosystem stability.